Predictive Maintenance for Electric Micro-Mobility Using Public Sensor Datasets: Battery State-of-Health Forecasting, Fault Diagnosis, and Tyre-Pressure–Aware Energy Consumption Modelling

Urban last-mile logistics is increasingly served by light electric micro-mobility vehicles, whose high-variability duty cycles accelerate heterogeneous degradation and intermittent faults. This study presents a reproducible, data-centric workflow for predictive maintenance using public sensor datasets and addresses 3 complementary tasks: (i) battery ageing prognostics via State-of-Health (SoH) trajectory forecasting, (ii) multi-class fault diagnosis from propulsion/thermal telemetry, and (iii) tyre-pressure–aware energy-consumption modelling. Battery ageing was analysed using the NASA Prognostics Center of Excellence dataset of 4 Li-ion 18650 cells (B0005, B0006, B0007, B0018). After cleaning and smoothing charge-phase records, an LSTM model (50 hidden units, 35 epochs, batch size 256, shuffle=False) tracked measured capacity fade with stable convergence; removing normalization produced a small reported degradation (average error increase of 0.3%). A degree-2 polynomial regression baseline captured the global decay trend but generalized less effectively (test R² = 0.6861, test MAE = 0.0274). For fault diagnosis on the “New Energy Vehicles Diagnosis” dataset, Random Forest achieved the highest test performance (accuracy 0.899, macro-F1 0.900, macro-AUC 0.985), followed by SVM (RBF) and logistic regression. For energy consumption regression including tyre pressure, linear regression showed consistent generalization (test R² = 0.9474, test MSE = 0.2528) under nominal pressure conditions (≈ 28–35 psi). Overall, the results indicate that task-appropriate model selection and disciplined preprocessing can yield reliable, interpretable predictive signals for maintenance planning in micro-mobility contexts.